Model fusion research aims to aggregate the knowledge of multiple individual models to enhance performance by combining their weights. In this work, we study the inverse problem: investigating whether model fusion can be used to reduce unwanted knowledge. We investigate the effects of model fusion in three scenarios: the learning of shortcuts, social biases, and memorization of training data in fine-tuned language models. Through experiments covering classification and generation tasks, our analysis highlights that shared knowledge among models is enhanced during model fusion, while unshared knowledge is usually forgotten. Based on this observation, we demonstrate the potential of model fusion as a debiasing tool and showcase its efficacy in addressing privacy concerns associated with language models.

Despite their high predictive accuracies, current machine learning systems often exhibit systematic biases stemming from annotation artifacts or insufficient support for certain classes in the dataset. Recent work proposes automatic methods for identifying and explaining systematic biases using keywords. We introduce DISCERN, a framework for interpreting systematic biases in text classifiers using language explanations. DISCERN iteratively generates precise natural language descriptions of systematic errors by employing an interactive loop between two large language models. Finally, we use the descriptions to improve classifiers by augmenting classifier training sets with synthetically generated instances or annotated examples via active learning. On three text-classification datasets, we demonstrate that language explanations from our framework induce consistent performance improvements that go beyond what is achievable with exemplars of systematic bias. Finally, in human evaluations, we show that users can interpret systematic biases more effectively (by over 25% relative) and efficiently when described through language explanations as opposed to cluster exemplars.

While much research has explored enhancing the reasoning capabilities of large language models (LLMs) in the last few years, there is a gap in understanding the alignment of these models with social values and norms. We introduce the task of judging social acceptance. Social acceptance requires models to judge and rationalize the acceptability of people’s actions in social situations. For example, is it socially acceptable for a neighbor to ask others in the community to keep their pets indoors at night? We find that LLMs’ understanding of social acceptance is often misaligned with human consensus. To alleviate this, we introduce SocialGaze, a multi-step prompting framework, in which a language model verbalizes a social situation from multiple perspectives before forming a judgment. Our experiments demonstrate that the SocialGaze approach improves the alignment with human judgments by up to 11 F1 points with the GPT-3.5 model. We also identify biases and correlations in LLMs in assigning blame that is related to features such as the gender (males are significantly more likely to be judged unfairly) and age (LLMs are more aligned with humans for older narrators).

A hallmark of human intelligence is the ability to learn new concepts purely from language. Several recent approaches have explored training machine learning models via natural language supervision. However, these approaches fall short in leveraging linguistic quantifiers (such as ‘always’ or ‘rarely’) and mimicking humans in compositionally learning complex tasks. Here, we present LaSQuE, a method that can learn zero-shot classifiers from language explanations by using three new strategies - (1) modeling the semantics of linguistic quantifiers in explanations (including exploiting ordinal strength relationships, such as ‘always’ > ‘likely’), (2) aggregating information from multiple explanations using an attention-based mechanism, and (3) model training via curriculum learning. With these strategies, LaSQuE outperforms prior work, showing an absolute gain of up to 7% in generalizing to unseen real-world classification tasks.

Psychology research has long explored aspects of human personality like extroversion, agreeableness and emotional stability, three of the personality traits that make up the ‘Big Five’. Categorizations like the ‘Big Five’ are commonly used to assess and diagnose personality types. In this work, we explore whether text generated from large language models exhibits consistency in it’s perceived ‘Big Five’ personality traits. For example, is a language model such as GPT2 likely to respond in a consistent way if asked to go out to a party? We also show that when exposed to different types of contexts (such as personality descriptions, or answers to diagnostic questions about personality traits), language models such as BERT and GPT2 consistently identify and mirror personality markers in those contexts. This behavior illustrates an ability to be manipulated in a predictable way (with correlations up to 0.84 between intended and realized changes in personality traits), and frames them as tools for controlling personas in applications such as dialog systems. We contribute two data-sets of personality descriptions of humans subjects.

Recent approaches have explored language- guided classifiers capable of classifying examples from novel tasks when provided with task-specific natural language explanations, instructions or prompts (Sanh et al., 2022; R. Menon et al., 2022). While these classifiers can generalize in zero-shot settings, their task performance often varies substantially between different language explanations in unpredictable ways (Lu et al., 2022; Gonen et al., 2022). Also, current approaches fail to leverage unlabeled examples that may be available in many scenarios. Here, we introduce TALC, a framework that uses data programming to adapt a language-guided classifier for a new task during inference when provided with explanations from multiple teachers and unlabeled test examples. Our results show that TALC consistently outperforms a competitive baseline from prior work by an impressive 9.3% (relative improvement). Further, we demonstrate the robustness of TALC to variations in the quality and quantity of provided explanations, highlighting its potential in scenarios where learning from multiple teachers or a crowd is involved. Our code is available at: https://github.com/WeiKangda/TALC.git.

Real-world domain experts (e.g., doctors) rarely annotate only a decision label in their day-to-day workflow without providing explanations. Yet, existing low-resource learning techniques, such as Active Learning (AL), that aim to support human annotators mostly focus on the label while neglecting the natural language explanation of a data point. This work proposes a novel AL architecture to support experts’ real-world need for label and explanation annotations in low-resource scenarios. Our AL architecture leverages an explanation-generation model to produce explanations guided by human explanations, a prediction model that utilizes generated explanations toward prediction faithfully, and a novel data diversity-based AL sampling strategy that benefits from the explanation annotations. Automated and human evaluations demonstrate the effectiveness of incorporating explanations into AL sampling and the improved human annotation efficiency and trustworthiness with our AL architecture. Additional ablation studies illustrate the potential of our AL architecture for transfer learning, generalizability, and integration with large language models (LLMs). While LLMs exhibit exceptional explanation-generation capabilities for relatively simple tasks, their effectiveness in complex real-world tasks warrants further in-depth study.

Generalized quantifiers (e.g., few, most) are used to indicate the proportions predicates satisfy (for example, some apples are red). One way to interpret quantifier semantics is to explicitly bind these satisfactions with percentage scopes (e.g., 30%-40% of apples are red). This approach can be helpful for tasks like logic formalization and surface-form quantitative reasoning (Gordon and Schubert, 2010; Roy et al., 2015). However, it remains unclear if recent foundation models (Bommasani et al., 2021) possess this ability due to the absence of direct training signals. To explore this, we introduce QuRe, a crowd-sourced dataset of human-annotated generalized quantifiers in Wikipedia sentences featuring percentage-equipped predicates. We explore quantifier comprehension using PRESQUE, a framework that combines natural language inference and the Rational Speech Acts framework. Experimental results on the HVD dataset (Herbelot and Vecchi, 2015) and QuRe demonstrate PRESQUE’s superiority over a literal listener baseline, showing a 20% relative improvement in F1 in predicting percentage scopes for quantifiers, even with no additional training.

Understanding the internal reasoning behind the predictions of machine learning systems is increasingly vital, given their rising adoption and acceptance. While previous approaches, such as LIME generate algorithmic explanations by attributing importance to input features for individual examples, recent research indicates that practitioners prefer examining language explanations that explain sub-groups of examples (Lakkaraju et al., 2022). In this paper, we introduce MaNtLE, a model-agnostic natural language explainer that analyzes a set of classifier predictions and generates faithful natural language explanations of classifier rationale for structured classification tasks. MaNtLE uses multi-task training on thousands of synthetic classification tasks to generate faithful explanations. Our experiments indicate that, on average, MaNtLE-generated explanations are at least 11% more faithful compared to LIME and Anchors explanations across three tasks. Human evaluations demonstrate that users can better predict model behavior using explanations from MaNtLE compared to other techniques.

While large language models have shown exciting progress on several NLP benchmarks, evaluating their ability for complex analogical reasoning remains under-explored. Here, we introduce a high-quality crowdsourced dataset of narratives for employing proverbs in context as a benchmark for abstract language understanding. The dataset provides fine-grained annotation of aligned spans between proverbs and narratives, and contains minimal lexical overlaps between narratives and proverbs, ensuring that models need to go beyond surface-level reasoning to succeed. We explore three tasks: (1) proverb recommendation and alignment prediction, (2) narrative generation for a given proverb and topic, and (3) identifying narratives with similar motifs. Our experiments show that neural language models struggle on these tasks compared to humans, and these tasks pose multiple learning challenges.

Supervised learning has traditionally focused on inductive learning by observing labeled examples of a task. In contrast, a hallmark of human intelligence is the ability to learn new concepts purely from language. Here, we explore training zero-shot classifiers for structured data purely from language. For this, we introduce CLUES, a benchmark for Classifier Learning Using natural language ExplanationS, consisting of a range of classification tasks over structured data along with natural language supervision in the form of explanations. CLUES consists of 36 real-world and 144 synthetic classification tasks. It contains crowdsourced explanations describing real-world tasks from multiple teachers and programmatically generated explanations for the synthetic tasks. To model the influence of explanations in classifying an example, we develop ExEnt, an entailment-based model that learns classifiers using explanations. ExEnt generalizes up to 18% better (relative) on novel tasks than a baseline that does not use explanations. We delineate key challenges for automated learning from explanations, addressing which can lead to progress on CLUES in the future. Code and datasets are available at: https://clues-benchmark.github.io.

Item Response Theory (IRT) has been extensively used to numerically characterize question difficulty and discrimination for human subjects in domains including cognitive psychology and education (Primi et al., 2014; Downing, 2003). More recently, IRT has been used to similarly characterize item difficulty and discrimination for natural language models across various datasets (Lalor et al., 2019; Vania et al., 2021; Rodriguez et al., 2021). In this work, we explore predictive models for directly estimating and explaining these traits for natural language questions in a question-answering context. We use HotpotQA for illustration. Our experiments show that it is possible to predict both difficulty and discrimination parameters for new questions, and these traits are correlated with features of questions, answers, and associated contexts. Our findings can have significant implications for the creation of new datasets and tests on the one hand and strategies such as active learning and curriculum learning on the other.

Large language models (LMs) have rapidly become a mainstay in Natural Language Processing. These models are known to acquire rich linguistic knowledge from training on large amounts of text. In this paper, we investigate if pre-training on text also confers these models with helpful ‘inductive biases’ for non-linguistic reasoning. On a set of 19 diverse non-linguistic tasks involving quantitative computations, recognizing regular expressions and reasoning over strings. We find that pretrained models significantly outperform comparable non-pretrained neural models. This remains true also in experiments with training non-pretrained models with fewer parameters to account for model regularization effects. We further explore the effect of text domain on LMs by pretraining models from text from different domains and provenances. Our experiments surprisingly reveal that the positive effects of pre-training persist even when pretraining on multi-lingual text or computer code, and even for text generated from synthetic languages. Our findings suggest a hithertho unexplored deep connection between pre-training and inductive learning abilities of language models

Transformer-based models have shown promising performance in numerous NLP tasks. However, recent work has shown the limitation of such models in showing compositional generalization, which requires models to generalize to novel compositions of known concepts. In this work, we explore two strategies for compositional generalization on the task of kinship prediction from stories, (1) data augmentation and (2) predicting and using intermediate structured representation (in form of kinship graphs). Our experiments show that data augmentation boosts generalization performance by around 20% on average relative to a baseline model from prior work not using these strategies. However, predicting and using intermediate kinship graphs leads to a deterioration in the generalization of kinship prediction by around 50% on average relative to models that only leverage data augmentation.

Existing approaches for the Table-to-Text task suffer from issues such as missing information, hallucination and repetition. Many approaches to this problem use Reinforcement Learning (RL), which maximizes a single manually defined reward, such as BLEU. In this work, we instead pose the Table-to-Text task as Inverse Reinforcement Learning (IRL) problem. We explore using multiple interpretable unsupervised reward components that are combined linearly to form a composite reward function. The composite reward function and the description generator are learned jointly. We find that IRL outperforms strong RL baselines marginally. We further study the generalization of learned IRL rewards in scenarios involving domain adaptation. Our experiments reveal significant challenges in using IRL for this task.

Mapping natural language instructions to programs that computers can process is a fundamental challenge. Existing approaches focus on likelihood-based training or using reinforcement learning to fine-tune models based on a single reward. In this paper, we pose program generation from language as Inverse Reinforcement Learning. We introduce several interpretable reward components and jointly learn (1) a reward function that linearly combines them, and (2) a policy for program generation. Fine-tuning with our approach achieves significantly better performance than competitive methods using Reinforcement Learning (RL). On the VirtualHome framework, we get improvements of up to 9.0% on the Longest Common Subsequence metric and 14.7% on recall-based metrics over previous work on this framework (Puig et al., 2018). The approach is data-efficient, showing larger gains in performance in the low-data regime. Generated programs are also preferred by human evaluators over an RL-based approach, and rated higher on relevance, completeness, and human-likeness.

Contextual representations learned by language models can often encode undesirable attributes, like demographic associations of the users, while being trained for an unrelated target task. We aim to scrub such undesirable attributes and learn fair representations while maintaining performance on the target task. In this paper, we present an adversarial learning framework “Adversarial Scrubber” (AdS), to debias contextual representations. We perform theoretical analysis to show that our framework converges without leaking demographic information under certain conditions. We extend previous evaluation techniques by evaluating debiasing performance using Minimum Description Length (MDL) probing. Experimental evaluations on 8 datasets show that AdS generates representations with minimal information about demographic attributes while being maximally informative about the target task.

Recently, pre-trained language models (LMs) have achieved strong performance when fine-tuned on difficult benchmarks like SuperGLUE. However, performance can suffer when there are very few labeled examples available for fine-tuning. Pattern Exploiting Training (PET) is a recent approach that leverages patterns for few-shot learning. However, PET uses task-specific unlabeled data. In this paper, we focus on few-shot learning without any unlabeled data and introduce ADAPET, which modifies PET’s objective to provide denser supervision during fine-tuning. As a result, ADAPET outperforms PET on SuperGLUE without any task-specific unlabeled data.

Online forums such as ChangeMyView have been explored to research aspects of persuasion and argumentative quality in language. While previous research has focused on arguments between a view-holder and a persuader, we explore the premise that apart from the merits of arguments, persuasion is influenced by the ambient social community. We hypothesize that comments from the rest of the community can either affirm the original view or implicitly exert pressure to change it. We develop a structured model to capture the ambient community’s sentiment towards the discussion and its effect on persuasion. Our experiments show that social features themselves are significantly predictive of persuasion (even without looking at the actual content of discussion), with performance comparable to some earlier approaches that use content features. Combining community and content features leads to overall performance of 78.5% on the persuasion prediction task. Our analyses suggest that the effect of social pressure is comparable to the difference between persuasive and non-persuasive language strategies in driving persuasion and that social pressure might be a causal factor for persuasion.

We explore learning web-based tasks from a human teacher through natural language explanations and a single demonstration. Our approach investigates a new direction for semantic parsing that models explaining a demonstration in a context, rather than mapping explanations to demonstrations. By leveraging the idea of inverse semantics from program synthesis to reason backwards from observed demonstrations, we ensure that all considered interpretations are consistent with executable actions in any context, thus simplifying the problem of search over logical forms. We present a dataset of explanations paired with demonstrations for web-based tasks. Our methods show better task completion rates than a supervised semantic parsing baseline (40% relative improvement on average), and are competitive with simple exploration-and-demonstration based methods, while requiring no exploration of the environment. In learning to align explanations with demonstrations, basic properties of natural language syntax emerge as learned behavior. This is an interesting example of pragmatic language acquisition without any linguistic annotation.

Recent work by Clark et al. (2020) shows that transformers can act as “soft theorem provers” by answering questions over explicitly provided knowledge in natural language. In our work, we take a step closer to emulating formal theorem provers, by proposing PRover, an interpretable transformer-based model that jointly answers binary questions over rule-bases and generates the corresponding proofs. Our model learns to predict nodes and edges corresponding to proof graphs in an efficient constrained training paradigm. During inference, a valid proof, satisfying a set of global constraints is generated. We conduct experiments on synthetic, hand-authored, and human-paraphrased rule-bases to show promising results for QA and proof generation, with strong generalization performance. First, PRover generates proofs with an accuracy of 87%, while retaining or improving performance on the QA task, compared to RuleTakers (up to 6% improvement on zero-shot evaluation). Second, when trained on questions requiring lower depths of reasoning, it generalizes significantly better to higher depths (up to 15% improvement). Third, PRover obtains near perfect QA accuracy of 98% using only 40% of the training data. However, generating proofs for questions requiring higher depths of reasoning becomes challenging, and the accuracy drops to 65% for “depth 5”, indicating significant scope for future work.

Natural language has recently been explored as a new medium of supervision for training machine learning models. Here, we explore learning classification tasks using language in a conversational setting – where the automated learner does not simply receive language input from a teacher, but can proactively engage the teacher by asking questions. We present a reinforcement learning framework, where the learner’s actions correspond to question types and the reward for asking a question is based on how the teacher’s response changes performance of the resulting machine learning model on the learning task. In this framework, learning good question-asking strategies corresponds to asking sequences of questions that maximize the cumulative (discounted) reward, and hence quickly lead to effective classifiers. Empirical analysis across three domains shows that learned question-asking strategies expedite classifier training by asking appropriate questions at different points in the learning process. The approach allows learning classifiers from a blend of strategies, including learning from observations, explanations and clarifications.

People can identify correspondences between narratives in everyday life. For example, an analogy with the Cinderella story may be made in describing the unexpected success of an underdog in seemingly different stories. We present a new task and dataset for story understanding: identifying instances of similar narratives from a collection of narrative texts. We present an initial approach for this problem, which finds correspondences between narratives in terms of plot events, and resemblances between characters and their social relationships. Our approach yields an 8% absolute improvement in performance over a competitive information-retrieval baseline on a novel dataset of plot summaries of 577 movie remakes from Wikipedia.

Humans can efficiently learn new concepts using language. We present a framework through which a set of explanations of a concept can be used to learn a classifier without access to any labeled examples. We use semantic parsing to map explanations to probabilistic assertions grounded in latent class labels and observed attributes of unlabeled data, and leverage the differential semantics of linguistic quantifiers (e.g., ‘usually’ vs ‘always’) to drive model training. Experiments on three domains show that the learned classifiers outperform previous approaches for learning with limited data, and are comparable with fully supervised classifiers trained from a small number of labeled examples.

We present a generative probabilistic model of documents as sequences of sentences, and show that inference in it can lead to extraction of long-range latent discourse structure from a collection of documents. The approach is based on embedding sequences of sentences from longer texts into a 2- or 3-D spatial grids, in which one or two coordinates model smooth topic transitions, while the third captures the sequential nature of the modeled text. A significant advantage of our approach is that the learned models are naturally visualizable and interpretable, as semantic similarity and sequential structure are modeled along orthogonal directions in the grid. We show that the method is effective in capturing discourse structures in narrative text across multiple genres, including biographies, stories, and newswire reports. In particular, our method outperforms or is competitive with state-of-the-art generative approaches on tasks such as predicting the outcome of a story, and sentence ordering.

We present LIA, an intelligent personal assistant that can be programmed using natural language. Our system demonstrates multiple competencies towards learning from human-like interactions. These include the ability to be taught reusable conditional procedures, the ability to be taught new knowledge about the world (concepts in an ontology) and the ability to be taught how to ground that knowledge in a set of sensors and effectors. Building such a system highlights design questions regarding the overall architecture that such an agent should have, as well as questions about parsing and grounding language in situational contexts. We outline key properties of this architecture, and demonstrate a prototype that embodies them in the form of a personal assistant on an Android device.

Natural language constitutes a predominant medium for much of human learning and pedagogy. We consider the problem of concept learning from natural language explanations, and a small number of labeled examples of the concept. For example, in learning the concept of a phishing email, one might say ‘this is a phishing email because it asks for your bank account number’. Solving this problem involves both learning to interpret open ended natural language statements, and learning the concept itself. We present a joint model for (1) language interpretation (semantic parsing) and (2) concept learning (classification) that does not require labeling statements with logical forms. Instead, the model prefers discriminative interpretations of statements in context of observable features of the data as a weak signal for parsing. On a dataset of email-related concepts, our approach yields across-the-board improvements in classification performance, with a 30% relative improvement in F1 score over competitive methods in the low data regime.